Early prediction of Type 2 Diabetes Mellitus (T2DM) is crucial for effective prevention and management. While Machine Learning (ML) has shown promise on structured hospital data, little attention has been given to primary care data. This study investigates the performance of several state-of-the-art ML models (including Logistic Regression, SVM, KNN, Decision Tree, Random Forest, XGBoost) and the recent TabPFN foundation model on a novel real-world multi-center dataset (FIMMG-6GP) collected from six general practitioners (GPs). We evaluate models using two validation strategies: stratified Five-Fold cross-validation (5F-CV) to assess within-distribution performance, and Leave-One-GP-Out (L1GPO-CV) to capture cross-practice variability. XGBoost consistently outperformed all models in terms of AUC (93.70% in 5F-CV; 93.10% in L1GPO-CV) and sensitivity, while also providing clinically interpretable insights via SHAP analysis. In contrast, TabPFN achieved high overall scores but showed poor sensitivity in detecting T2DM cases. Our findings underscore the value of the proposed cross-practice evaluation in developing trustworthy ML-based decision support systems for real-world clinical settings, and particularly for primary care.